What problem does it solve? Research experiments often produce results that look convincing but lack verified evidence: missing logs, unverified citations, threshold misses, or mismatched methods. This Skill acts as an independent evidence gate between experiment and analysis, ensuring no claim is promoted without typed, verifiable proof. ## Core Features & Use Cases - Four-lane review rubric: Assesses reproducibility/integrity, specification and metric compliance, citation authenticity, and method-code-log alignment using only pass, fail, needs-human, or not-applicable verdicts. - Evidence Pack verification: Runs run_review.sh <run-id> against the experiment's manifest and contract, validating hashes, logs, metrics, and external long-task (tmux) attempt chains before any claim promotion. - Typed review artifacts: Produces a Markdown review plus a ds-lite.review-result.v1 JSON sidecar, links them into the project state graph, and blocks analysis nodes until a passing review is active. - Use Case: After an experiment run finishes, invoke this Skill to verify the Evidence Pack, check that cited sources are authentic, and receive a gate decision (pass/fail/needs-human) with the smallest required follow-up before writing any analysis. ## Quick Start Use ds-lite-review to check whether the latest experiment run has enough valid evidence to support an analysis.
